Deep Variational Bayesian Modeling of Haze Degradation Process

TL;DR

This paper tackles single image dehazing by explicitly modeling uncertainties in the haze degradation process through a variational Bayesian framework. It treats the haze-free image $z$ and the transmission map $\tau$ as latent variables, parameterizing their posteriors with a dehazing network (D-Net) and a transmission network (T-Net) under the physical haze model $I = J \odot t + A(1-t)$. The authors derive an ELBO-based objective that couples the two networks via the likelihood term while regularizing with priors on $z$ and $\tau$, and they show model-agnostic integration with existing dehazing networks yields consistent performance gains across synthetic and real datasets, plus improvements in downstream tasks like object detection. The framework is scalable and allows inference with the dehazing network alone, avoiding added inference overhead. Overall, the work provides a principled, adaptable method to quantify and leverage uncertainty in haze removal, improving robustness and practical applicability.

Abstract

Relying on the representation power of neural networks, most recent works have often neglected several factors involved in haze degradation, such as transmission (the amount of light reaching an observer from a scene over distance) and atmospheric light. These factors are generally unknown, making dehazing problems ill-posed and creating inherent uncertainties. To account for such uncertainties and factors involved in haze degradation, we introduce a variational Bayesian framework for single image dehazing. We propose to take not only a clean image and but also transmission map as latent variables, the posterior distributions of which are parameterized by corresponding neural networks: dehazing and transmission networks, respectively. Based on a physical model for haze degradation, our variational Bayesian framework leads to a new objective function that encourages the cooperation between them, facilitating the joint training of and thereby boosting the performance of each other. In our framework, a dehazing network can estimate a clean image independently of a transmission map estimation during inference, introducing no overhead. Furthermore, our model-agnostic framework can be seamlessly incorporated with other existing dehazing networks, greatly enhancing the performance consistently across datasets and models.

Figures (9)

• Figure 1: Our variatonal Bayesian framework is model-agnostic, and consistently improves the performance of existing dehazing neural networks across different benchmark datasets (SOTS li2018benchmarking, Haze4K haze4k and NH-Haze Ancuti_2020_CVPR_Workshops) in terms of PSNR and SSIM values. Upward-right movement of the star indicates better restoration.
• Figure 2: The architecture of proposed variational network. Blue solid lines represent forward process and red dotted lines denote gradient flow in back-propagation. Note that our D-Net and T-Net are not depending on specific network architectures. In addition, we can only employ D-Net to output the haze-free image in the inference stage, hence no additional overhead during inference.
• Figure 3: User study results.
• Figure 4: Visual comparisons between the baseline models and our enhanced models on the SOTS dataset li2018benchmarking. (a) Input hazy image. (b) GCANet chen2018gated. (c) GCANet + Ours. (d) FFA-Net qin2020ffa. (e) FFA-Net + Ours. (f) DehazeFormer-B song2022vision. (g) DehazeFormer-B + Ours. Best viewed on high-resolution display.
• Figure 5: Visual comparisons of image dehazing methods on Fattal evaluation set Fattal2014. (a) Hazy input image. (b) GCANet chen2018gated. (c) GCANet + Ours. (d) FFA-Net qin2020ffa. (e) FFA-Net + Ours. (f) DehazeFormer song2022vision. (g) DehazeFormer + Ours.